claim
active
claim:reward-hacking-representations-reduce-to-more-generic-cheating-representationsReward hacking representations reduce to more generic cheating representations
Interpretation of logit lens, exemplars, steering.
Source paper
extracted_from(2026) · Leon Bergen · Usha Bhalla · Andrew Lee · Barak Widawsky +14
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (4)
finding
- Probes encode cheating concept.
- Generalization to cheating concepts.
- Steering along probe makes exam stories feature cheating, shortcuts, and SWE/evaluation themessupportsBehavioral effect of probe.
- Causal role of probe.
Questions (1)
question
- Open scientific question motivating the work.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Speculation.
- Opening question.
- Exploiting unintended high-reward behaviors; tested in combination with alignment faking
- Distinct from sandbagging view of awareness.
- Central difficulty of learned reward models.
- Simple DoM vectors coherently represent reward hacking and are generalizable and interpretableclaim0.779Central claim.
- Reward hacking generalizes to broader deceptive behaviors even when core misalignment score is 0%
- Fourteen-fold increase.